Before this we tried hard to optimize for the sysmem case by emitting
the events for the renderpass inline and only discarding them in the
gmem case. However this won't work if we need to emit the
render_pass_start event after the binning IB, because we don't know
whether the binning IB will be emitted until the RP end time. The old
system also required a ton of confusing code to keep track of the
start/end pointers with suspending and resuming renderpasses. All of
that goes away if we use a separate u_trace for renderpass events and
just copy it to the main trace buffer for sysmem.
With this, the previous method of using the space between trace_rp_start
and trace_rp_drawcalls_start to keep track of the renderpass_start event
and disable it when emitting a split dynamic renderpass at submit time
doesn't work anymore. Just move trace_renderpass_start() to
tu_cmd_render() time after the draw calls have been recorded, which is
now safe to do.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35751>
Connor Abbott
630380349b